Underwater Object Detection Overhaul: Infrastructure and Generalizability

Abstract

The Michigan Robosub Student Project Team develops an underwater autonomous submarine to complete a variety of tasks in an international competition. The submarine uses a camera and machine learning to detect objects needed to complete these tasks. These objects can be things like buoys or gates. Since the underwater environments we run in are not always consistent, sometimes the machine learning model struggles to accurately recognize these objects. As a project, I have overhauled to team’s machine learning infrastructure. In the old system, there is a very complex process of generating labels and running commands on a remote computer, which I hope to streamline. The team also needs a established data collection method. This will ensure we always add images to our test set. We also need a new evaluation metric for our models, as there is currently no data-backed way to say one model performs better than another. Finally, I want to research and experiment how to increase our model’s performance in a range of environments. I will try methods such as training on certain subsets of the data and color correction, along with other ideas found through research, to try and increase our model’s object detection abilities.